JP2012027617A - Pattern identification device, pattern identification method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make it possible to easily check in which state a plurality of registered images are.SOLUTION: A pattern identification device has registration means for registering images in a dictionary and confirmation means for confirming the dictionary images already registered in the dictionary. The confirmation means displays the relationships of the plurality of dictionary images. The relationship of the dictionary images is positional relationship on a characteristic space used for pattern identification or a positional relationship on a projection space on which the characteristic amount used for pattern identification is projected.

Description

  The present invention relates to a pattern identification device, a pattern identification method, and a program, and more particularly to a technique suitable for use in identifying a dictionary image.

  As an identification technique in pattern recognition, typically, as a technique for identifying that an object that is a subject in image data is the same as an object that is a subject in another image, for example, an individual's face is identified. There is face recognition technology. As a face identification technique, for example, there is a method as described in Non-Patent Document 1. This method is an algorithm that enables face registration / additional learning in real time by replacing an individual identification problem by a face with a two-class identification problem of a feature class called a differential face.

  For example, in the face identification using a well-known support vector machine (SVM), in order to identify the faces of n persons, the faces of registered persons and other faces are identified. n SVM discriminators are required. When registering a person's face, SVM learning is also required. SVM learning requires a large amount of face data of a person to be registered and already registered persons and other persons, and it takes a lot of calculation time. Met.

  However, according to the method described in Non-Patent Document 1, additional learning can be substantially eliminated by replacing the problem of personal identification with the following two classes of identification problems. These two classes are a variation feature class (extra-personal class) between images of the same person and a variation feature class (intra-personal class) such as illumination variation, facial expression / orientation, etc., and a different person image. . Assuming that the distribution of the two classes is constant regardless of a specific individual, the classifier is configured by reducing the individual face identification problem to the two class identification problem. A large number of images are prepared in advance, and the discriminator is learned so as to identify the variation feature class between the same persons and the variation feature class between different persons. Thereby, the new registrant only needs to hold the result of extracting the facial image or necessary features. When discriminating, the difference feature is extracted from the two images, and the discriminator determines whether the person is the same person or a different person. This eliminates the need for learning such as SVM at the time of personal face registration, and registration can be performed in real time.

  As described above, in an apparatus that specifically identifies an object in an image, more specifically, a person's face, fluctuation between a registration pattern and an authentication pattern is a factor that degrades identification performance. Is mentioned. That is, there are fluctuations in the object (person's face) that is the identification target, specifically lighting conditions, orientation / posture, hiding by other objects, fluctuation due to facial expressions, and the like. When the above-described fluctuations increase between the registration pattern and the authentication pattern, the identification performance is greatly reduced. As a solution to such a problem, for example, in the technique described in Patent Document 1, when acquiring an image for registration, a fitness level indicating whether it is suitable for registration is displayed on a screen, and a material for judgment to the user is obtained. Is given.

JP 2009-48447 A Japanese Patent No. 3078166 Japanese Patent Laid-Open No. 2002-8032

Baback Moghaddam, Beyond Eigenfaces: Probabilistic Matching for Face Recognition (M.I.T Media Laboratory Perceptual Computing Section Technical Report No.433), ProbabilisticVisual Learning for Object Representation (IEEE Transactions on PatternAnalysis and Machine Intelligence, Vol. 19, No. 7, JULY 1997)

  However, the user has no way of confirming how the actually registered image affects the identification performance, and has only to acquire and register a registered image recommended by the device. In addition, when the number of registered images is limited, information for determining which registered image should be deleted is not sufficiently given.

  The present invention has been made in view of the above-described problems, and an object of the present invention is to make it possible to easily check the state of a plurality of registered images.

  The pattern identification apparatus of the present invention includes registration means for registering an image in a dictionary, and confirmation means for confirming a dictionary image already registered by the registration means, wherein the confirmation means includes a plurality of dictionary images. The relationship is displayed.

  According to the present invention, the user can easily confirm the state of the plurality of registered dictionary images, and can select a more appropriate registered image.

It is a block diagram which shows the function structural example of the pattern identification apparatus which concerns on embodiment. It is a flowchart which shows an example of the whole process sequence of a pattern identification device. It is a block diagram which shows the detailed functional structural example of an input image identification part. It is a flowchart which shows an example of the identification processing procedure by an input image identification part. It is a block diagram which shows the detailed functional structural example of a dictionary image registration part. It is a block diagram which shows the detailed functional structural example of a dictionary image confirmation part. It is a flowchart which shows an example of the process sequence by a dictionary image relationship calculating part. It is a schematic diagram which shows the example of a display by a dictionary image display part. It is a block diagram which shows the detailed functional structural example of a dictionary image registration part. It is a block diagram which shows the detailed functional structural example of a dictionary image confirmation part. It is a flowchart which shows an example of the process sequence by a dictionary image group evaluation part.

(First embodiment)
Hereinafter, a first embodiment of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a block diagram illustrating a configuration example of a pattern identification device 100 according to the present embodiment.
As shown in FIG. 1, the pattern identification device 100 includes an imaging optical system 1, an imaging unit 2, an imaging control unit 3, an image recording unit 4, a dictionary image registration unit 5, an input image identification unit 6, and a dictionary image confirmation unit. 7 is provided. Further, an external output unit 8 for outputting the identification result, an operation unit 9 for the user to perform various operations, and a connection bus 10 for performing control / data connection of each component are further provided. The dictionary image registration unit 5 and the input image identification unit 6 may typically be a dedicated circuit (ASIC) and a processor (reconfigurable processor, DSP, etc.), respectively. Further, it may exist as a program executed in a single dedicated circuit and general-purpose circuit (PC CPU).

  The imaging optical system 1 includes an optical lens having a zoom mechanism. Further, a drive mechanism in the pan / tilt axis direction may be provided. As the image sensor of the imaging unit 2, a CCD or a CMOS image sensor is typically used. A predetermined video signal (for example, a signal obtained by sub-sampling or block reading) is output from the imaging unit 2 as image data by a read control signal from a sensor drive circuit (not shown). The imaging control unit 3 determines the timing at which actual shooting is performed based on an instruction from the photographer (viewing angle adjustment instruction, shutter pressing, etc.) and information from the dictionary image registration unit 5 or the input image identification unit 6. Control.

  The image recording unit 4 is configured by a semiconductor memory or the like, holds image data transferred from the imaging unit 2, and at a predetermined timing in response to a request from the dictionary image registration unit 5 or the input image identification unit 6. Transfer image data. The dictionary image registration unit 5 extracts information of an object to be identified from the image data, and records and holds it. Note that a more detailed configuration in the dictionary image registration unit 5 and specific contents of processing actually performed will be described later.

  The input image identification unit 6 identifies an object in the image data based on the image data and the data acquired from the dictionary image registration unit 5. In addition, regarding the input image identification part 6, a specific structure and the detail of the process performed are mentioned later. The dictionary image confirmation unit 7 is used by the user to confirm the contents of the dictionary image. The dictionary image registration unit 5 can also be used to confirm registration candidate images to be registered. Details of processing and display performed in the dictionary image confirmation unit 7 will be described later.

  The external output unit 8 is typically a monitor such as a CRT or a TFT liquid crystal, and displays the image data acquired from the imaging unit 2 and the image recording unit 4, or the dictionary image registration unit 5 and the input image as image data. The output results of the identification unit 6 and the dictionary image confirmation unit 7 are superimposed and displayed. Also, the output results of the dictionary image registration unit 5, the input image identification unit 6, and the dictionary image confirmation unit 7 may be output as electronic data to an external memory or the like.

  The operation unit 9 allows the user to perform various operations such as mode setting for performing a dictionary image registration operation and operations for a registered dictionary image. The connection bus 10 is a bus for performing control and data connection between the above components. Hereinafter, the pattern identification means that a difference between individual patterns (for example, a difference between persons as individuals) is determined. On the other hand, the detection of the pattern means that an object that falls within the same category is determined without distinguishing between individuals (for example, a face is detected without distinguishing between individuals).

  FIG. 2 is a flowchart showing an example of the overall processing procedure of the pattern identification apparatus 100 according to this embodiment. Hereinafter, an actual process in which the pattern identification apparatus 100 identifies an input pattern will be described with reference to FIG. In the following, a case where the pattern to be identified is an object in an image, more specifically, a human face will be described, but it goes without saying that the present invention is not limited to this.

  First, the input image identification unit 6 acquires image data from the image recording unit 4 (step S200). Subsequently, target pattern detection, specifically, human face detection processing is performed on the acquired image data (step S201). A known technique may be used as a method for detecting a human face from an image. For example, a technique proposed in Patent Document 2 or Patent Document 3 can be used. Even when a face other than a person's face is detected, it is possible to detect the desired object as training data using the technique described above.

  Next, it is determined whether or not the target pattern exists in the input image (step S202). If the result of this determination is that there is no target pattern, the process is terminated. On the other hand, if the target pattern exists in the input image as a result of the determination in step S202, it is determined whether or not the registration mode is set (step S203). This registration mode can be set in advance by performing a predetermined operation on the operation unit 9 by the user. If the result of this determination is registration mode, the dictionary image registration unit 5 performs dictionary image registration processing (step S204). The specific contents of the dictionary image registration process will be described in detail later.

  On the other hand, if the result of determination in step S203 is not registration mode, the input image identification unit 6 performs input image identification processing (step S205). Specific contents of this input image identification processing will also be described later. Finally, the result of the dictionary image registration process or the input image identification process is output to the external output unit 8 (step S206).

Next, the input image identification process will be described. FIG. 3 is a block diagram illustrating a detailed functional configuration example of the input image identification unit 6.
As shown in FIG. 3, the input image identification unit 6 includes an input image identification data generation unit 21, a dictionary image data acquisition unit 22, and an input image identification calculation unit 23.

  The input image identification data generation unit 21 extracts information necessary for identifying the target object from the image data acquired from the image recording unit 4. The dictionary image data acquisition unit 22 acquires dictionary data necessary for identification of the input image from the dictionary image registration unit 5. The input image identification calculation unit 23 performs input image identification processing from the identification data acquired from the input image identification data generation unit 21 and the dictionary data acquired from the dictionary image data acquisition unit 22.

FIG. 4 is a flowchart illustrating an example of an identification processing procedure performed by the input image identification unit 6.
First, dictionary data is acquired from the dictionary image registration unit 5 (step S400). This dictionary data is typically a normalized image obtained by cutting out a target object from an input image, correcting the inclination, and expanding or reducing the object to a predetermined size. Alternatively, the same processing as the input image identification data generation processing described later may be performed in advance, and the feature data extracted may be used as dictionary data.

  Next, image data is acquired from the image recording unit 4 (step S401). Then, input image identification data generation processing is performed (step S402). Here, when the dictionary data is an image, identification data generation processing is performed not only on the input image but also on the dictionary data. In the input image identification data generation processing, first, a target object (typically a human face) is cut out from input image data, and after normalizing in a predetermined size and rotation direction, a feature amount is extracted. . Here, there is a method that simply uses the brightness of the image as it is, but as a feature that is more robust against fluctuations, a feature that has been subjected to local binary pattern or incremental code is used for the brightness. May be. In addition, a feature vector whose dimension is reduced by a principal component analysis method (PCA) or an independent component analysis method (ICA) may be used as a new feature value.

  Next, matching processing is performed on the dictionary data acquired in step S400 and the identification data acquired in step S401 (step S403). The matching process typically performs a correlation operation between feature vectors extracted from the dictionary data and the input data. When there are a plurality of dictionary data, the brute force is performed for all of them. The correlation calculation may be a simple inner product or a correlation coefficient. In addition, as the output of the matching process, when matching with registered data (dictionary data) is output in binary (0 or 1), the normalized output value is output as a (0-1 real value) likelihood. A case may be considered. Further, when there are a plurality (a plurality of) registered patterns (registrants), the likelihood may be output for each registered pattern (registrant). Alternatively, only the result for the registered pattern that matches the best may be output. In addition, the likelihood for the class to which the registered pattern belongs may be output instead of the likelihood for the registered pattern. That is, in the case of a person, the likelihood for the person ID (name) is output instead of the result for each registered face image.

Next, the dictionary registration process will be described. FIG. 5 is a block diagram illustrating a detailed functional configuration example of the dictionary image registration unit 5.
As shown in FIG. 5, the dictionary image registration unit 5 includes a dictionary image data generation unit 31, a dictionary image data holding unit 32, a dictionary image data selection unit 33, and a registration candidate image confirmation unit 34.

  The dictionary image data generation unit 31 generates dictionary image data necessary for identifying an individual pattern from the image data acquired from the image recording unit 4. For example, as described in Non-Patent Document 1, when a two-class problem of intra-class and extra-class is determined, typically, a human face image may be used as dictionary image data. Further, the face image data of a person detected by the face detection process may be stored in the dictionary image data holding unit 32 after normalizing the size and direction (in-plane rotation direction). At this time, label information set by the user through the operation unit 9 or the like is also recorded on the dictionary image. In this way, it is possible to reduce the amount of dictionary data by holding only the data necessary for identification rather than the image data itself. When performing the discrimination calculation by taking the vector correlation of the partial area of the pattern, it is sufficient to cut out only the partial area in advance.

  Further, the dictionary image data generation unit 31 can perform processing on an image designated by the user via the dictionary image confirmation unit 7 described later. As described above, necessary information is appropriately extracted from the image, subjected to predetermined conversion described later, and then stored in the dictionary image data holding unit 32 as a feature vector for identifying an object.

  The dictionary image data selection unit 33 reads necessary dictionary image data from the dictionary image data holding unit 32 and transfers the dictionary image data to the input image identification unit 6 in response to the request from the input image identification unit 6 described above. The registration candidate image confirmation unit 34 includes a display device for a user to confirm dictionary image data input to the dictionary image data generation unit 31 and an interface for operation. An example of processing by the registration candidate image confirmation unit 34 will be described again in the description of the dictionary image confirmation unit 7 described later.

Next, processing performed in the dictionary image confirmation unit 7 will be described. FIG. 6 is a block diagram illustrating a detailed functional configuration example of the dictionary image confirmation unit 7.
As shown in FIG. 6, the dictionary image confirmation unit 7 includes a dictionary image relationship calculation unit 41, a dictionary image relationship display unit 42, and a dictionary image relationship operation unit 43.

  The dictionary image relationship calculation unit 41 acquires a plurality of dictionary image data from the dictionary image data holding unit 32, and performs a process for calculating the relationship between the dictionary images as a numerical value. The contents of this process will be described in detail later. The dictionary image relationship display unit 42 is a display unit for displaying the output result of the dictionary image relationship calculation unit 41. This display method will also be described in detail later. The dictionary image related operation unit 43 is an operation device for the user to perform various operations on the dictionary image group based on the result of the dictionary image related display unit 42.

Next, the content of the process performed by the dictionary image relation calculating unit 41 will be described. FIG. 7 is a flowchart illustrating an example of a processing procedure performed by the dictionary image relation calculation unit 41.
First, dictionary image data is acquired from the dictionary image data holding unit 32 (step S700). Subsequently, a feature amount is acquired from the acquired dictionary image data (step S701). Here, as the feature amount, the feature amount generated by the input image identification data generation unit 21 may be used as it is. In order to reduce the calculation load, typically, simpler feature amounts may be extracted, for example, luminance values are used as feature vectors as they are.

  Next, the extracted feature quantity is extracted as a feature vector (step S702). Subsequently, the dimension of the extracted feature vector is reduced (step S703). Here, in the dimension reduction process, a method similar to the process performed in the input image identification data generation unit 21 may be employed. Typically, techniques such as principal component analysis (PCA) and independent component analysis (ICA) are used. The number of dimensions after the reduction may be the same as or different from the value of the input image identification data generation unit 21. Then, it is determined whether or not the processing has been completed for all dictionary image data (step S704). As a result of the determination, if all are completed, the process proceeds to step S705, and if there is unprocessed data, the process returns to step S701.

  Next, a relative coordinate calculation process is performed (step S705). In this process, the relative relationship between the dictionary images is calculated using the feature vector after dimension reduction of all dictionary image data acquired in step S703. Specifically, this is performed as follows. First, correlation values are calculated for all combinations of dictionary images, and the similarity between the dictionary images is digitized. Then, the similarity between all dictionary images is output as a table to a dictionary image relationship display unit 42 described later. In this way, the relationship between the registered image and other registered images (similarity) can be visualized by the dictionary image relationship display unit 42.

  Further, instead of the correlation value, the feature vector can be reduced in dimensions from two to three dimensions so that the positional relationship can be illustrated. In this case, in the feature vector dimension reduction process in step S703, the number of dimensions after reduction may be set to 2 to 3. When using PCA, typically, two or three base vectors having a high degree of contribution are selected and projected onto the space spanned by the base vectors as the positional relationship of the dictionary image. Further, a multidimensional scale construction method (MDS) may be used. According to MDS, feature vectors can be reduced in order so that similar ones are close and different ones are arranged far away, and the feature amount space can be displayed as a representative.

  In this manner, the positional relationship of the dictionary image can be visualized by reducing the feature vector to one, two, or three dimensions and outputting it to the dictionary image relationship display unit 42. In other words, the user can check the dictionary image in almost the same space (on the projection space) as the feature space where the actual identification is performed.

  Next, the contents of processing performed in the dictionary image relation display unit 42 will be described. The dictionary image relationship display unit 42 is for visualizing and displaying the output result of the dictionary image relationship calculation unit 41, and typically includes a monitor device such as an LCD.

FIG. 8 is a schematic diagram illustrating a display example by the dictionary image relation display unit 42.
In FIG. 8, the dictionary image is mapped to a two-dimensional graph. That is, the dictionary image relation display unit 42 reduces the feature vector of the dictionary image to two dimensions, and displays the result on a graph. When displaying an icon obtained by reducing the dictionary image on the graph at the time of display, the positional relationship between the plurality of dictionary images can be confirmed more intuitively. A label defined by the user (in the case of a human face, a personal ID) may be displayed under the icon display.

  It is expected that the dictionary images displayed through the above processing will be displayed at close positions if they have the same label, and those that have different labels will be displayed at distant positions. Is done. If the dictionary image with the same label is at a relatively distant position, it is presumed that it is not preferable for identification due to some condition. On the other hand, even when dictionary images with different labels are close to each other, it is not a preferable state for identification. Thus, by displaying the relationship between a plurality of dictionary images, the user can intuitively understand how the registered dictionary images look.

  Further, the dictionary image confirmation unit 7 and the registration candidate image confirmation unit 34 can be linked. For example, the input image input to the dictionary image registration unit 5 is input to the dictionary image confirmation unit 7 via the registration candidate image confirmation unit 34, and the dictionary image relation calculation unit 41 is the same as other dictionary image data. Process. The result of processing the registration candidate image can be displayed on the dictionary image relationship display unit 42 to show the relationship with other dictionary images. By doing so, the user can intuitively understand the relationship between the image to be registered and the dictionary image that has already been registered. Furthermore, it is more effective to perform the above processing in real time. In other words, the subject to be registered can be photographed while changing the angle of view and the illumination conditions, and it is possible to immediately confirm the relationship with the already registered dictionary image group.

  As another example of linking the dictionary image confirmation unit 7 and the registration candidate image confirmation unit 34, the user may be advised based on the state of the registration candidate image. For example, the same processing as in the above cooperation example is performed to display the registration candidate image together with the dictionary image registered in the dictionary image confirmation unit 7, and display information about what the registration candidate image will be and how the display position changes. To do. As described above, the state of the dictionary image can be confirmed by the positional relationship displayed on the dictionary image relationship display unit 42. In other words, it is desirable that the dictionary images with the same label are displayed at close positions and the dictionary images with different labels are displayed at distant positions. The input image that is the registration candidate is advised so as to be in the above state with respect to other dictionary images. Specifically, advice is given on how to obtain a better dictionary image, such as the orientation of the face as the subject, lighting conditions, facial expressions, and the like.

  How to set the content of the advice can be obtained by collecting and learning a sufficient amount of typical subject data in advance. That is, it can be realized by obtaining from the sample data what kind of locus the change in the face direction and the illumination condition will be in the display space of the dictionary image relation display unit. In this way, the user can easily set the shooting conditions without trial and error as to how a preferred dictionary image can be registered.

(Second Embodiment)
In the present embodiment, the configurations of the dictionary image registration unit and the dictionary image confirmation unit are partially different from those of the first embodiment. Hereinafter, this embodiment will be specifically described. In addition, in order to avoid duplication, the same part as 1st Embodiment is abbreviate | omitted in the following description. Since the overall configuration of the pattern identification apparatus according to this embodiment is the same as that of FIG. 1 described in the first embodiment, the description thereof is omitted.

FIG. 9 is a block diagram illustrating a detailed functional configuration example of the dictionary image registration unit 5.
As shown in FIG. 9b, the dictionary image registration unit 5 includes a dictionary image data generation unit 51, a dictionary image data holding unit 52, a dictionary image data selection unit 53, a registration candidate image confirmation unit 54, and a dictionary image deletion unit 55. ing. Compared to the first embodiment, the point that the dictionary image deletion unit 55 is added is different from the part of the processing contents of the dictionary image data selection unit 53.

  The dictionary image data selection unit 53 not only selects and transfers the dictionary image data held by the dictionary image data holding unit 52 in response to a request from the input image identification unit 6 but also a user through the operation unit 9 and the like. One or more dictionary images can be selected in response to a request from. The dictionary image deletion unit 55 deletes the dictionary image selected by the dictionary image data selection unit 53 from the dictionary image data holding unit 52.

FIG. 10 is a block diagram illustrating a detailed functional configuration example of the dictionary image confirmation unit 7.
As shown in FIG. 10, the dictionary image confirmation unit 7 includes a dictionary image relationship calculation unit 61, a dictionary image relationship display unit 62, a dictionary image relationship operation unit 63, a dictionary image group evaluation unit 64, and a dictionary image relationship input unit 65. I have. Compared with the first embodiment, a dictionary image group evaluation unit 64 and a dictionary image relation input unit 65 are added.

  The dictionary image group evaluation unit 64 evaluates the overall state of all registered dictionary images using the dictionary image data and the calculation result of the dictionary image relation calculation unit 61. The specific contents of the processing performed by the dictionary image group evaluation unit 64 will be described in detail later. The dictionary image relationship input unit 65 performs a process for the user to input the relationship between the label information of a plurality of dictionary images. Specific contents of the processing performed by the dictionary image relation input unit 65 will be described later.

Next, processing contents of the dictionary image group evaluation unit 64 will be described. FIG. 11 is a flowchart illustrating an example of a processing procedure performed by the dictionary image group evaluation unit 64.
First, all registered dictionary image data is acquired from the dictionary image data holding unit 52 (step S1100). Subsequently, data representing the relationship between all dictionary image data is acquired from the dictionary image relationship calculation unit 61 (step S1101). Then, the attribute value of the dictionary image data acquired in step S1100 is acquired (step S1102). Here, the attribute values of the dictionary image data typically include lighting conditions, the size and orientation of the target object, and the like. When the target object is a person's face, it may include a facial expression, gender, age, and the like. These attributes of the dictionary image data may be detected directly from the dictionary image data, and the illumination conditions may be estimated from camera parameters recorded at the time of imaging.

  Next, the attributes of the dictionary image data acquired in step S1102 are totaled for each label information of the dictionary image (step S1103). That is, the variation of the attribute is measured for the attribute of the dictionary image data to which the same label information is assigned. Next, the dictionary image relation data acquired in step S1101 and the attribute variation in the same label dictionary image acquired in step S1103 are comprehensively evaluated to calculate an evaluation value of the dictionary image group (step S1104).

  The evaluation value can be calculated based on, for example, the number of labels in the dictionary image, the number of dictionary images belonging to the same label, and attribute variations. In general, if the number of labels in the dictionary image is small, the number of classes to be identified is small, so that a high value can be given as the evaluation value of the dictionary image group. In addition, since the number of dictionary image data belonging to the same label is large and various attributes are included, the identification performance is further improved, so that the evaluation value of the evaluation image group can be increased.

  A method that uses the feature amount for evaluation instead of the number and attributes of the dictionary image is also conceivable. For example, it is possible to measure the distribution of feature quantities belonging to different labels from the feature quantities belonging to the same label, and add the degree of separation to the evaluation value. The degree of feature vector separation may be obtained using the following equation (1).

  The function d is a function for obtaining a distance between two vectors, and typically calculates a Euclidean distance between the two vectors. The average of the distances between the dictionary images belonging to the labels C1 and C2 as described above is defined as the degree of separation D. Generally, the higher the degree of separation between dictionary images belonging to different labels, the higher the identification performance is expected. That is, it is possible to obtain the evaluation value of the dictionary image group by obtaining the separation degree for all combinations of labels and averaging them.

  In addition, instead of the distance between the feature amounts as shown in Expression (1), a ratio between the variance between the feature amounts belonging to the same label and the variance between the feature amount groups belonging to different labels can be used. For example, an amount as shown in the following formula (2) is obtained.

  The denominator of J in equation (2) represents the variance between feature vectors belonging to the same label, and the numerator represents the variance of feature vectors belonging to different labels. Generally, it is considered that the discrimination performance is improved when the feature vectors belonging to different labels are separated and the feature vectors belonging to the same label are collected. That is, it is desirable that the variance between feature vectors belonging to different labels is large and the variance of feature vectors belonging to the same label is small. Therefore, the evaluation value of a dictionary image group can be made high, so that J is large.

  As described above, by calculating the evaluation value of the dictionary image group and displaying it on the dictionary image confirmation unit, the user can easily confirm the state of the dictionary image registered in the device. Is possible.

  Next, the dictionary image relation input unit 65 will be described. The dictionary image relationship input unit 65 has a function for a user to input a relationship between labels of dictionary images. Here, the relationship between the labels of the dictionary image is a relationship between the persons when the target object is a person's face and the label is the name of the person, for example. More specifically, a blood relationship such as “parent and child” and “brother” is input. It should be noted that a relationship other than a blood relationship such as “friend” or “colleague” may be input.

  As described above, the relationship between registered dictionary image labels input by the user can be used for the evaluation of the dictionary image group evaluation unit 64. For example, the relationship between different labels can be used for weighting when calculating the average degree of separation of dictionary images belonging to different labels in the evaluation operation of the dictionary image group evaluation unit. In general, it is considered that there is a tendency that the degree of separation of feature vectors differs depending on whether a parent or child or a sibling is related. That is, it is expected that the feature vectors of the persons having a parent-child / brother relationship will have a lower degree of separation than that of others even if different labels are attached. Simply taking the average of the degree of separation, the latter has a higher evaluation value in the dictionary image group including the related-related labels and the dictionary image group other than that. In addition, when the dictionary image is additionally registered, the latter is more advantageous for the evaluation value of the dictionary image group when adding a related person and when adding another person. In order to reduce such a phenomenon, it is preferable to lower the weight of the feature vector separation degree between different labels when there is a blood relationship.

(Other embodiments)
The present invention can also be realized by executing the following processing. That is, software (program) that realizes the functions of the above-described embodiments is supplied to a system or apparatus via a network or various storage media, and a computer (or CPU, MPU, or the like) of the system or apparatus reads the program. It is a process to be executed.

5 dictionary image registration unit, 6 input image identification unit, 7 dictionary image confirmation unit

Claims (5)

Registration means for registering images in a dictionary;
Confirmation means for confirming a dictionary image already registered by the registration means,
The pattern identifying apparatus, wherein the confirmation unit displays a relationship between a plurality of dictionary images.   The pattern identification apparatus according to claim 1, wherein the relationship between the dictionary images is a positional relationship on a feature space used for pattern identification.   The pattern identification apparatus according to claim 1, wherein the relationship between the dictionary images is a positional relationship on a projection space obtained by projecting a feature amount used for pattern identification. A registration process for registering images in a dictionary;
A confirmation step of confirming a dictionary image already registered in the registration step,
In the confirmation step, a relationship between a plurality of dictionary images is displayed.   The program for making a computer perform each process of the pattern identification method of Claim 4.

Images